[axisrow]

Batch 6: reproduce upstream TUI/skills fixes on macOS — 4031a93, 75f07d6

TDD verification per the /fix pipeline: red on the pre-fix tree → green on main under both runners (bun test + vitest/Node) → mutation check (StrykerJS) → lint. Last batch of the wave; follows the merged #55/#56/#57 and open #58 (batches 4–5).

Upstream CI only runs Linux; 4031a93 is security-relevant (imported shell commands were applied without showing them), 75f07d6 is a path-traversal fix (session ids naming files outside the skills folder).

Fix	Red fact (pre-fix tree)	Green bun	Green vitest	Mutation	Commit
75f07d6 — skills log stays inside its folder	all 7 traversal ids (x/../../outside, ../outside, a/b, .., empty, with space, a\b) get a log path; a hook with a traversal id writes outside the skills folder @ f51332a	9/9	9/9	skills.ts/hook-handler.ts subprocess-invisible (see below)	c72505e
4031a93 — imported shell commands shown and confirmed	both confirm cases fail: an import adding a shell command applies immediately (onApply on the first Enter, no confirmation screen), Replace All and Merge @ 610dcc9	3/3	12/12	pure functions of the fix: 0 survivors (see below)	aeaca1d

Harness notes

• skills-path tests run the hook in a subprocess (bun-bundled probe executed by node, HOME in a temp directory): the hook's writes land in the temp home, never in the live one.
• import-confirm tests drive the real ink component headless (MockTtyStream + waitFor, upstream's own pattern).

Survivors triage (fix lines)

• escapeControlCharacters / getImportedCommands / getCustomCommands (pure functions added by the fix): zero surviving mutants.
• ImportPreviewDialog.tsx confirm-screen wiring (~25 survivors): all display-only — item labels, descriptions, the confirmation prompt text and the selection-change display handling. None inverts the apply/cancel choice, which the main test pins end to end: first Enter lands on Cancel (no apply), returning to the preview, then ArrowDown + Enter applies exactly once with onApply('replace').
• skills.ts / hook-handler.ts report uncovered rather than surviving: the subprocess harness's coverage is invisible to the mutation runner. The red/green pair above is the regression proof (same triage class as usage-fetch in batch 5 and powerline in batch 4).

Notes

• Full suite on this branch: bun 3595 tests, 0 failures in the recorded run (the known macOS-only LFS case did not reproduce this time; it remains pre-existing and flaky).
• An unrelated untracked docs/review-remaining-prs-2026-10-12.md exists in the working tree of the main checkout — left alone deliberately.

Commits

1. c72505e — test(skills): reproduce upstream 75f07d6 on macOS
2. aeaca1d — test(tui): reproduce upstream 4031a93 on macOS

🤖 Generated with Claude Code

[axisrow]

Review: PR #59 — approved

Test-only PR reproducing upstream fixes 4031a93 (import: shell commands shown and confirmed) and 75f07d6 (skills log path traversal) on macOS. Two new test files, no source changes. Approving — no changes requested.

Verified during review:

• Red claims hold. At f51332a, src/utils/skills.ts has no SAFE_SESSION_ID (traversal session ids got a log path); at 610dcc9, ImportPreviewDialog.tsx has no pendingMode confirm flow. Both test files would fail on those trees and pass on this base (origin/main 6979c23 already contains both fixes).
• APIs match the source. getSkillsFilePath returns null exactly for ids failing [A-Za-z0-9_-]+ (covers all 7 parametrized ids incl. empty, space, backslash); handleHookInput returns before mkdirSync when the path is null, so the escapedFile/ccstatuslineDir absence assertions are meaningful; the dialog test pins the full flow end to end — escaped command shown in full (\u001b[2J literal), first Enter lands on the Cancel-preselected confirm screen with onApply uncalled, Enter on Cancel returns to preview, then ArrowDown+Enter applies exactly onApply('replace'); and a no-commands import still applies immediately.
• Hermetic. The skills probe runs in a subprocess with HOME/USERPROFILE pointed at a temp dir (node honors HOME), so no writes reach the real home; bun build + node per test is the documented harness.
• Convention match. MockTtyStream/waitFor/down-arrow (\u001B[B) patterns mirror the sibling ink tests; all imports are used; assertions are not tautological.

Non-blocking notes:

• getOutput() in the new dialog test does not strip ANSI, unlike every sibling test in the directory. It works because every asserted string is a contiguous single-Text run (selected labels are one chalk-wrapped run, the command line and the yellow warning are single uncolored/colored texts), but it is a latent fragility if ink's rendering ever splits runs; stripping would make it uniform with siblings.
• The MockTtyStream harness is copy-pasted rather than shared — consistent with existing per-file convention, though #56 showed the dedup path exists if the ink harness ever grows a shared helper too.